JPH0291698A - Sound encoding and decoding system - Google Patents

Abstract

PURPOSE:To reproduce a satisfactory audio signal by using a multiple pulse train obtained by pitch forecasting and a code book indicating a narrow- amplitude sound source signal together as a sound source signal. CONSTITUTION:A discrete audio signal is inputted from an input terminal 500, and a pitch parameter and a spectrum parameter are obtained by a pitch parameter calculating circuit 515 and a spectrum parameter calculating circuit 520. A sound source pulse calculating circuit 580 obtains a pitch-forecasted multiple pulse train, and a narrow-amplitude sound source calculating circuit 620 obtains a pitch-forecasted code book. A multiplexer 630 combines and outputs the multiple pulse train and the code book. This signal passes a demultiplexer 710, and the multiple pulse train, the code book, and the pitch parameter are used to restore a sound source signal, and further, the spectrum parameter is used to reproduce a synthesized audio signal. Thus, the audio signal is satisfactorily reproduced.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an audio encoding/decoding method for efficiently encoding and decoding audio signals at a low bit rate.

(Prior Art) A multi-pulse encoding method is known as a method for transmitting audio signals at a low bit rate, for example, about 16 Kb/S or less. These represent the sound source signal as a combination of multiple pulses (multipulse), represent the characteristics of the vocal tract using a digital finoscillator, and transmit information on the sound source pulse and filter coefficients after determining them at fixed time intervals (frames). are doing. For details of this method, see, for example, Arasek
l, Ozawa, Ono, and 0chla1
Document by Mr. (hereinafter referred to as “Reference 1”) “Multi-pu
lse Excited 5peechCoder B
a5ed on Maximum Cross-cor
relationshipSearch Algorithm-
(IEEE Global Telecommunica
tion conference, 23.3+1
983). This method outputs a good audio signal after decoding by separately expressing the vocal tract information and the sound source signal, and by using a combination of multiple pulse trains (multipulse) as a means of expressing the sound source signal. . The basic concept of obtaining a multi-pulse train representing a sound source signal will be explained using FIG.

An audio signal in sections divided into frames is inputted from an input terminal 800 in the figure. Spectral parameters determined from the audio signal of the current frame are input to the synthesis filter circuit 820. An initial sound source multi-pulse train is generated in the sound source calculation circuit 810 and inputted to the synthesis filter 820 to obtain a synthesized speech waveform as an output. A subtracter 840 subtracts the synthesized speech waveform from the input signal.

This result is input to a weighting circuit 850, and the weighting error power in the current frame is obtained as an output. This weighting is performed by the sound source calculation circuit 810 so as to minimize the error power.
Find the amplitude and position of the sound source multi-pulse train at .

On the other hand, regarding the pitch prediction multi-pulse method that improves the sound quality of the method of Document 1 by performing pitch prediction using pitch parameters representing the fine structure of the pitch,
Since it is explained in the specification of No. 8-139022 (Document 2), the explanation is omitted here.

(Problem to be solved by the invention) However, in the conventional method of Document 1, the sound quality was good when the pill rate was sufficiently high and the number of sound source pulses was sufficient, but as the bit rate was lowered, the sound quality deteriorated. It was declining. In particular, in the conventional method, in the case of an input signal with a high pitch frequency, for example, when a female voice is input,
There was a drawback that the reproduced sound deteriorated. This means that when the pitch frequency is high, many pitch waveforms are included in the pulse calculation frame, and in order to reproduce this pitch waveform well, it is necessary to reproduce the pitch waveform well compared to the case of a speaker with a low pitch frequency. , because it requires a larger number of multi-pulses. Therefore, for this reason, it has been difficult to significantly reduce the transmission bit rate, that is, to significantly reduce the number of pulses within one frame, without degrading sound quality.

On the other hand, in the conventional method of Document 2, although pitch prediction is performed using pitch parameters based on the correlation for each pitch, multipulses and pitch prediction are used regardless of whether the sound source signal is a large amplitude sound source signal or a small amplitude sound source signal. represents the sound source signal. It is thought that a large-amplitude sound source signal has a high correlation for each pitch, but a small-amplitude sound source signal is considered to have a low correlation. In order to further improve the sound quality by this method, the role of a small-amplitude multi-pulse train or a small-amplitude sound source signal among the multi-pulse trains representing the sound source signal is even more important. Is this in particular? - Important for tonal audio signals. In the conventional method, only a set number of pulses are determined and transmitted in descending order of amplitude as a multi-pulse train representing the sound source signal.

Therefore, in the conventional array, due to the upper limit of the amount of information set in advance, it is not possible to obtain a sufficient number of small amplitude pulses, and the approximation of the sound source signal is insufficient (there is a limit in terms of the quality of the reproduced sound). It was not possible to improve the sound quality above, and this was especially noticeable when the bit rate was low.

SUMMARY OF THE INVENTION An object of the present invention is to provide an audio encoding/decoding method that is capable of reproducing audio better than before even when the bit rate is high or low.

(Means for Solving the Problems) The audio encoding/decoding method of the present invention inputs a discrete audio signal, and expresses a pitch parameter representing the fine structure of the pitch of the audio signal and a spectrum of the audio signal. A spectral parameter is obtained, and is expressed and transmitted using a multi-pulse train obtained by pitch-predicting the audio signal of the audio signal and a codebook obtained by pitch-predicting, and the codebook, the multi-pulse train, and the pitch parameter are expressed and transmitted. The spectral parameters are used to restore the sound source signal, and the spectral parameters are used to output a synthesized audio signal that satisfactorily represents the audio signal.

(Function) In the pitch predictive multipulse encoding method of Document 2, the present invention uses a pitch predictive multipulse and a code block to express the sound source signal more effectively than the conventional method with a small amount of transmitted information. It is characterized by being expressed as .

The operation of the present invention will be explained using FIG. 2. The upper part of FIG. 2 is a block diagram of a circuit that uses pitch prediction to obtain multipulses to represent the sound source signal. An audio signal in sections divided into frames is inputted from an input terminal 300 in the figure.

The pitch parameter determined from the audio signal of the current frame is input to the pitch reproduction filter 315. Spectral parameters determined from the audio signal of the current frame are input to the spectral envelope filter circuit 320. An initial sound source multi-pulse train is generated in the multi-pulse sound source 310, and this is passed through the pitch reproduction filter 315.
A driving sound source signal is obtained by inputting the signal into the . By inputting the driving sound source signal to the spectral envelope filter 320, a synthesized speech waveform is obtained as an output. A subtracter 340 subtracts the synthesized speech waveform from the input signal. This result is input to the weighting circuit 350, and the weighting error power in the current frame is obtained as an output. Then, the amplitude and position of the sound source multi-pulse train in the multi-pulse sound source 310 are determined so as to minimize the error power of this weighting. Switch 380 is always connected to the upper side when obtaining a multi-pulse train. The switch 380 is connected to the lower side after determining the multi-pulse train, and outputs the determined multi-pulse train, the synthesized speech waveform x (n) obtained by the pitch recovery filter 315 and the spectral envelope filter circuit 320 to the subtracter 220.

The small-amplitude sound source calculation circuit 230 calculates a small-amplitude sound source signal that represents the characteristics of the small-amplitude component of the sound source signal for each of the small sections (for example, about 5 m5 ec) that are obtained by further dividing the frame section.

Here, this small amplitude sound source signal has almost random phase characteristics and is considered to be almost like a noise signal. In order to encode such signals very efficiently, it is necessary to create a food book (codebook) of multiple small amplitude sound source signals in advance.
It is possible to use a vector quantization method that prepares and encodes. For vector quantization, for example, R, M
, a paper by Mr. Gray (hereinafter referred to as Document 3) “vecto
r quantization for 5peech
The explanation is omitted here because it is well known in the "former ng and recognition".

The operation of the small amplitude sound source signal circuit 230 will be described below.

The subtracter 220 converts the reproduced signal x (n) to the original B voice waveform x
The residual signal e (n) resulting from the subtraction from (n) is temporally uniformly divided into small sections shorter than a frame by the time division circuit 240.

A plurality of codebooks (codebooks) 250 are prepared in advance, and for each small section divided by the time division circuit 240, one type from the codebook is input and the gain is calculated through the gain circuit 255. After matching, the pitch is recovered by a pitch recovery filter 267 using the same pitch parameters as the pitch recovery filter 315, and the synthesized residual signal τ(n ).

The subtractor 270 subtracts the composite residual signal ';e (n) from the input residual signal e (n). This result is manually input to the weighting circuit 280, and the weighting obtains the error power as an output. Here, 280 performs the same operation as 350. Then, the weighting is performed using the codebook 25 to minimize the error power.
Select the optimal index from among 0 and output that index.

Next, the small amplitude sound source signal is expressed using a codebook,
The actual method for selecting a food book will be explained below using a formula. The codebook selection method is calculated so as to minimize the error power E defined by the following equation.

E=Σ[(e(n)−g−;i(n)) *W(n
)]2(1) where e (n) is the input residual signal in FIG. 2, g is the gain multiplied by the gain circuit 255,
e (n) is a residual signal reproduced by one type of food book selected and a spectral envelope filter. lol
(n) shows the impulse response of a filter (the weighting in FIG. 2 is the same as the circuit 280), which is weighted with auditory sense in mind. When formula (+) is minimized with respect to g, it becomes the form of formula (2).

g=(Σe, (n) ew(n) ) / (Σew
(n)re(n)) (2>Here, ew(n)=e″(nL end W(n): n(n)
*p(n)l(n) tree W(n)(3a)ew(n):e
(n)*W(n) (3b)
The symbol * represents convolution. The denominator of equation (2) is τW (n
) autocorrelation (strictly speaking, covariance), the numerator is the cross-correlation of τW (n) and ew(n, ). Also, n in equation (3a)
(n) is a signal represented by one type of code selected in the codebook. Further, p(n) indicates the impulse response of the pitch recovery filter circuit 257, and h(n) indicates the impulse response of the spectral envelope filter circuit 260.

At this time, the error power E can be written as the following equation, so E=Σew(n)2−g・Σev(n);(n
) (4) The codebook that minimizes n n E may be selected to maximize the second term of equation (4), that is, to maximize g.

As a method for further significantly reducing the amount of calculation for selecting a codebook, the following configuration may be considered. A multi-pulse train representing a sound source signal is searched using cross-correlation. This calculation method is detailed in the above-mentioned documents 1 and 2, so the explanation is omitted here, but by using the cross-correlation function Φxh′ after calculating the pitch prediction multipulse train, the calculation Wffl can be significantly reduced compared to the method described above. After that, it is possible to select a codebook.

In the method shown below, the signal ew (
Since there is no need to reproduce n ), the amount of calculation can be significantly reduced while keeping the characteristics almost the same as in the method described above. The derivation method will be explained below. First, Φ.

’, ew(n) can be written as follows.

Φ. °:Σevi(n)hw(n) (5)e”w
(n) : n(n) small p(n) book hw(n)
(G) By substituting equation (6) into equation (2) and using equation (5), the following transformation is possible.

g= <ΣΦxh' ・n(n)) / (Rhh(0
) ”Rnn(0)) (9) Here, ΦXl+9 is the cross-correlation function after obtaining the multipulse train by pitch prediction, Rh+, (o) is the pitch recovery filter 257, the spectral envelope filter 260, and the weighting circuit 280. Rhh(0) is the power of the impulse response of the filter consisting of the cascade connection of Rhh(0) is the power of the signal n(n) represented by the code when one type of code is selected from the codebook 250. The numerator of formula (7) is Φxh′
and the signal n(n) represented by the selected code. As with the above-mentioned equation (2), the codebook that maximizes g in equation (7) can be selected.

Note that the food book 250 may be created by training in advance using the residual signal of the sound source after obtaining a predetermined number of large-amplitude pitch multipulse trains, or, for example, by training with a Gaussian It is also possible to create a plurality of noise signals having the statistical properties with variously changed phase characteristics. Regarding the latter method, M.
R.

Paper by Mr. 5hroeder and Mr. B, S, Ata1 (hereinafter referred to as Document 4): “Code-Excited 1i
near predlctron (CELP): h!
gh-quality 5peech at very
lov bitrates,” (Proc, 1.c
,A,s,s,P,vol,l,paper No.

25.1. I, March, +985).

The transmission information on the transmitting side is the amplitude and position of the pitch-predicted multipulse, the index and gain of the codebook of the small amplitude excitation signal, the pitch parameter, and the spectrum parameter.

(Embodiment) In FIG. 1 showing an embodiment of the present invention, an input terminal 50
A discrete audio signal x (n) from 0 is input. The time division circuit 510 temporally divides the input audio signal into various frames (for example, every 20 m5 ec). A pitch parameter calculation circuit 515 calculates a pitch parameter representing the pitch fine structure. The calculation method is described in the above document 2.
Use a method such as that shown in . ffi quantization evaporator 516 quantizes the determined pitch parameter. The dequantizer 518 dequantizes and outputs the quantized result. The spectral parameter calculation circuit 520 calculates spectral parameters representing the spectrum of the audio signal in the divided sections using the well-known LPC analysis method.

The obtained spectral parameters are quantized in a spectral parameter quantizer 525. The quantization method may be scalar quantization as shown in Japanese Patent Application No. 59-272435 (Reference 5) or vector quantization as shown in Reference 4.

The dequantizer 530 dequantizes and outputs the quantized result. A weighting circuit 540 weights the divided audio signals using the dequantized spectral parameters. For the weighting method, reference can be made to the weighting circuit 200 in Document 5. The impulse response calculation circuit 550 calculates an impulse response using the dequantized pitch parameter and the dequantized spectral parameter. For the specific method, reference can be made to the above-mentioned document 2. The autocorrelation function calculation circuit 560 calculates the autocorrelation function of the impulse response, and the sound source pulse calculation circuit 5
Output to 80. For the method of calculating the autocorrelation function, reference can be made to the autocorrelation function calculation circuit 180 in Document 2 mentioned above.

The cross-correlation function calculation circuit 570 calculates a cross-correlation function between the weighted signal and the impulse response and outputs it to the sound source pulse calculation circuit 580. For the specific method, reference can be made to the above-mentioned document 2.

The sound source pulse calculation circuit 580 calculates a predetermined number 81 (L 1) of multipulses by pitch prediction. Regarding the method of calculating a multi-pulse train, reference can be made to the sound source pulse calculation circuit 210 of the above-mentioned document 2. A quantizer 585 quantizes the excitation multipulse train and outputs a code. This output is dequantized by dequantizer 590 and regenerated into multiple pulses by pulse generator 600. The pitch reproduction filter 605 inputs the reproduced multi-pulse and the pitch parameter dequantized by the dequantizer 518, and outputs a pitch-reproduced sound source signal. By passing the sound source signal and the spectral parameters output from the inverse quantizer 530 through a spectral envelope filter 610, a synthesized speech signal '1(n) based on the sound source pulse is obtained.

The subtracter 615 subtracts the synthesized speech signal x(n) from the speech signal x(n), thereby producing a residual signal e(n
) to calculate the small amplitude sound source signal.

As explained in the operation section above, the small amplitude sound source calculation circuit 620 calculates a small amplitude sound source signal of a small section (for example, 5 m5ec) divided into sections shorter than a frame using a plurality of food books. represent.

The index and gain of the codebook representing the small amplitude excitation signal, the code obtained by quantizing the multi-pulse train that is the output of the quantizer 585, the code obtained by quantizing the pitch parameter that is the output of the quantizer 518, and the quantizer 525 The codes obtained by quantizing the spectral parameters which are the outputs of are respectively input to the multiplexer 630. The multiplexer 630 combines and outputs the above codes.

On the other hand, on the receiving side, the demultiplexer 710 separates and outputs the code of the multipulse train, the code of the pitch parameter, the code of the spectral parameter, and the code of the index and gain representing the small amplitude excitation signal. The source pulse decoder 720 decodes the amplitude and position of the multipulses. The spectral parameter decoder 750 includes the inverse quantizer 5 on the transmitting side.
It works the same as 30. The small amplitude excitation decoder 730 has the same codebook as the small amplitude excitation calculation circuit 820 on the transmitting side, and uses the received index to select and output a code representing the small amplitude excitation signal. Gain circuit 73
5 determines the amplitude of the small-amplitude excitation signal using the received gain sign when a code indicating that the current frame is in a consonant interval is received. Pitch parameter decoder 7
45 has the same output as the inverse quantizer 518 on the transmitting side. The pulse generator 725 generates a sound source signal based on the multi-pulse train. The pitch reproduction filter 755 inputs the weighted sound source signal and the decoded pitch parameter and reproduces a pitch-regenerated synthetic sound source signal. The adder 740 adds the pitch-reproduced sound source signal and the output signal of the gain circuit 735 to obtain a driving sound source signal, and drives the spectral envelope filter circuit 760. The spectral envelope filter circuit 760 uses the driving sound source signal and the decoded spectral parameters to obtain and output a synthesized speech waveform.

The configuration described above is only one configuration of the present invention, and various modifications are possible.

In order to further significantly reduce the amount of calculation required to obtain a small-amplitude sound source signal, as explained in equations (5) to (7) in the action section, the mutual It is possible to have a configuration in which a codebook is selected using correlation functions Φ and . By doing this, as mentioned in the section on the operation above, it is not necessary to reproduce the signal τw (n) when selecting the codebook, so the first
The amount of calculation can be significantly reduced compared to the configuration shown in the figure.

Furthermore, as a method for calculating multi-pulses, various well-known methods can be used in addition to the method shown in the above-mentioned document 1.

Further, as the spectral parameter, other well-known parameters (line spectrum pair, cepstrum, mel cepstrum, logarithmic cross-sectional area ratio, etc.) can also be used. Furthermore, as a quantization method for spectral parameters, vector quantization or the like can be used in addition to scalar quantization.

Regarding vector quantization, reference can be made to the above-mentioned document 3.

Further, although the frame length is fixed, it may be variable.

Also, a small amplitude sound source calculation circuit 620, a small amplitude sound source decoder 7
At step 30, at least one of a pitch period and a pitch gain as a pitch parameter used when selecting and decoding an optimal codebook may be recalculated from the residual signal e(ri) and used.

Also, the input signal can be converted into, for example, vowels, consonants (or voiced,
The food book may be used only for the consonant parts. In this case, the vowel and consonant parts can be determined using a well-known method such as parameters such as the power of the frame, the difference in power from the previous frame, the change in spectrum from the previous frame, and the pitch property. On the other hand, parameters such as pitch gain can be used to distinguish between child voices and unvoiced parts.

Further, for the consonant part, different food books may be created in advance according to the properties of the consonant (for example, plosiveness, fricativeness, etc.), and these may be used by switching.

Furthermore, it is also possible to determine the effectiveness of using the food book and use the codebook only for the parts that are effective, or to determine the effectiveness after determining the vowel and consonant parts. It is possible. To determine the effectiveness, the ratio of the power or RMS of the signal reproduced using the codebook to the power or RMS of the residual signal that is the output of the subtracter 615 in each section, the magnitude of the codebook gain, etc. can be used.

(Effects of the Invention) According to the present invention, compared to the conventional example, a codebook is used to represent not only a pitch-predicted multi-pulse train as a sound source signal but also a small-amplitude sound source signal that is more effective in improving sound quality. By using this, sound source information can be expressed efficiently. Therefore, even if the bit rate is the same as that of the conventional method, there is a great effect that a reproduced audio signal that is better than the conventional method can be obtained not only in the vowel portion but also in the consonant section. Furthermore, this effect becomes more pronounced when the bit rate is lowered.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech encoding/decoding method and apparatus according to the present invention, and FIG. 2 is a block diagram showing the operation of the present invention. FIG. 3 is a block diagram showing a conventional example of a multi-pulse train search method. In the figure, 240.510... time division circuit, 51
5... Pitch parameter calculation circuit, 520... Spectral parameter calculation circuit, 51B, 525, 585.
... Quantizer, 518.530.590 ... Inverse quantizer, 280.350.540.850 ... Weighting circuit, 550 ... Impulse response calculation circuit, 560 ...
- Autocorrelation function calculation circuit, 570... Cross correlation function calculation circuit, 580... Sound source pulse calculation circuit, 600.7
25...Pulse generator, 257.315.805.7
55... Pitch reproduction filter, 260.320.61
0.760...spectral envelope filter, 820...
-Synthesis filter circuit, 230.620...Small amplitude sound source calculation circuit, 630...Multiplexer, 710...
Demultiplexer, 720... Sound source pulse decoder, 7
30... Small amplitude sound source decoder, 740... Adder, 7
45... Pitch parameter decoder, 750... Spectrum parameter decoder, 770... Output terminal, 31
0...Multipulse sound source, 810...Sound source calculation circuit, 220,270.3401615.840...Subtractor, 255.735...Gain circuit, 380...Switch.

Claims (1)

[Claims] A discrete audio signal is input, a pitch parameter representing the pitch fine structure of the audio signal and a spectral parameter representing the spectrum of the audio signal are determined, and a multi-pulse train obtained by pitch predicting the sound source signal of the audio signal is calculated. Representing and transmitting using a codebook obtained by pitch prediction, restoring the sound source signal using the codebook, the multipulse train, and the pitch parameter, and satisfactorily representing the audio signal using the spectral parameter. A speech encoding/decoding method characterized by outputting a synthesized speech signal.